Auditory Based Feature Vectors for Speech Recognition Systems

Transcription

1 Auditory Based Feature Vectors for Speech Recognition Systems Dr. Waleed H. Abdulla Electrical & Computer Engineering Department The University of Auckland, New Zealand 1

2 Outlines Introduction ASR Systems and Signal Modelling The Human Ears Equivalent rectangular band (ERB) The Gammatone Filterbank (GTF) Speech Signal Analysis based on GTF Classification Evaluation Conclusions 2

3 Introduction Automatic speech recognition (ASR) is the process of converting an incoming acoustic signal to its corresponding stream of words. ASR systems can be: Speaker Dependent OR Speaker Independent Isolated Words OR Continuous Limited vocabulary OR Large vocabulary Resticted Domain OR Unrestricted Domain 3

4 Introduction The general paradigm of speech recognition systems comprises two main parts: front-end and back-end Signal Processing Part Statistical Modeling Part 4

5 Block diagram of the ASR systems ASR systems comprises: Speech Dataset Training Recognition s n Feature Extraction Recognition Phase O t Recognition W O t Training HMM Initial Training set Trainin, q t q t Training Phase 5

6 Speech Signal Processing Speech Signal Digital Filterbank Wavelets Fourier Transform Linear Prediction Power Estimation Mel Filterbank Fourier Transform Cepstrum Cepstrum PLP Coding Reflection Coefficients 6

7 MFCC & PLP Filterbanks MFCC Filterbank PLP Filterbank 7

8 Signal Modelling Feature Extraction Speech Signal s n Sampling Frequency Hz Preemphasis H(z)= z -1 Sampling at 9 ms rate Hanning Window 23 ms Time Domain Frequency Domain Power & 12 MFCC Normalised MFCC 12 Coefficients 36 & 72 ms Delta MFCC Delta-Delta MFCC Normalised Power 72 ms Delta Power Delta-Delta Power Feature Vectors Concatenation OR Streaming O t 8

9 The Structure of the Human Ears 9

10 Human Basilar Membrane 10

11 Cochlea characteristic frequency for different species In 1961, Don Greenwood developed a mathematical function relating the characteristic frequency, fc, at any location along the length of the cochlea to the distance, x, from the apex (Greenwood 1961). The function is: f c A(10 ax / L K) Where: A is a high frequency control constant L is the cochlea length in (mm) a is the slop factor K is the low frequency control constant 11

12 Reverse Correlation (Revcor) technique Revcore technique states that, for a linear system, it is possible to extract the system parameters by operations on stochastic input and output signals (de-boer and H. R. de Jongh 1978). The revcore function can be represented mathematically by the equation: g(t) t m e t cos( t) (a) (b) Amplitude e ff(t) Time (samples) Frequency 12

13 Critical band and equivalent rectangular bandwidth Critical band (CB) is the bandwidth of the human auditory filter at different characteristic frequencies positioned along the cochlea path. The bandwidth of the human auditory filter can be measured psycho-acoustically in masking experiments using a sine wave signal (single tone) and a broadband noise as a masker. Experiments show that sounds can be distinguished by ear only if they fall into different critical bands, and they practice the masking process on each other when they fall into the same critical band. H(f) 2 actual filter frequency 13

14 Equivalent rectangular bandwidth (ERB) The bandwidth of the actual auditory filter can be related to an equivalent rectangular bandwidth (ERB) filter that has a unit height and a bandwidth ERB. It passes the same power as the real filter does when subjected to a white noise input. ERB 0 H(f ) 2 df 14

15 Formulae for the ERB Various formulas have been derived for the ERB values: Zwicker 1961 ERB (1 1.4f c ) Glasberg and Moore 1990 ERB2 24.7(1 4.37f c ) Moore and Glasberg 1983 ERB3 6.23f c 93.39f c

16 Comparison of Different ERB Functions 16

17 General Formula for ERB ERB é m æ f c ö = + BW Q ê çè ø ë m min 1/ m ù úû Where f c is the centre frequency, Q is the ear quality factor, which is the ratio between the centre frequency and its corresponding filter bandwidth, BW min is the minimum bandwidth allowed, and m is the order. Lyon recommended the following parameters (Slaney 1988): Q = 8, BW min = 125 Hz, and m = 2 to produce ERB Ly ERB Ly é 2 æ f ö c = ç çè 8 ê ø ë 2 ù úû 17

18 General Formula for ERB Greenwood recommended: Q = 7.24, BW min = 22.85, m = 1 to form ERB Gr fc ERB Gr = Glassberg and Moore (Glasberg and Moore 1990) recommended Q = 9.26, BW min = 24.7, m = 1 to get ERB GM ERB GM fc = = 24.7( fc) 9.26 ERB GM is used in our approach as it approximates most of the other estimates. 18

19 Comparison Between Three ERB Definitions ERB ERB ERB Ly Gr GM 2 f c f c ( fc ) 19

20 Critical Band Number For a certain frequency, it represents the number of critical bands required until reaching that frequency. Let us consider the change in the critical-band number, z, as the frequency changes by df is given by: Dz 1 1 dz = df = df = df Df Df / Dz ERB( f ) z f c = ò 0 1 df ERB( f ) For ERB( f ) = 24.7( f ) fc 1 z= ò df = ln(4.37 f c + 1) 24.7( f ) 0 20

21 Gammatone Filters The impulse response of these filters ht nbt e wt ut n-1 -bt () = g(, ) cos( + f) () 21

22 ERB of Gammatone Filters 2 ERB = ò H ( f ) df 0 ( n -1)! 1 H( f) =. 2 é êë b + 4 p ( f - f ) c ù úû n /2-2( n-1) 2 ERB = 2 p( n -1)! b 2 ( 1)! [ n - ] For n = 4, ERB = b b = ERB 22

23 Number of Channels and the Overlapping Spacing z f f H = ò L 1 df ERB( f ) ERB f = + BWmin For m = 1, Q f H Q fh + QB z= ò df = Qln f + BQ f + QB f L L where B = BW min If the overlapping factor between the contiguous filters is then the number of channels, N, is related to z, as follows: z = N. v Q fh + QB 9.26 fh N =.ln = ln v f + QB v f L L For Q = 9.26 and B =

24 Gammatone Filterbank For a certain band f L f H with overlapping between filters N = 9.26 f H ln v f L For 1 n N f c ( n ) ( f H ) e vn 9.26 ERB ( n ) f ( n c ) 24

25 Characteristics of the GTF 25

26 Gammatone Filterbank Frequency response of a 30-channel filterbank, covering Hz band 26

27 Gammatone Filterbank Amplitude Time in samples Filter number is on the lower right corner Impulse responses of a 20-filters Gammatone filterbank. 27

28 Equal Loudness Contours This graph shows that the ear is not equally sensitive to all frequencies ISO recommendation R226 of equal loudness contours for pure tones and normal threshold of hearing for persons aged years. 28

29 Equal Loudness Preemphasis Filter The non-uniformity of the loudness sensing can be compensated for by a filter with the following transfer function E ( ) ( ) ( ).( ).( ) 29

30 Gammatone Filterbank Toward filter 20 Toward filter 1 Amplitude frequency responses of a 20-filters Gammatone filterbank after subjecting the filters to the equal loudness pre-emphasis filter. 30

31 Speech Signal GTF Frequency Analysis Speech Signal Frames Speech signal analysis of a spoken digit 9 using 30 Gammatone filters. (a) Spectra of the speech signal, (b) Log spectra of the speech signal. 31

32 Feature Extraction Paradigms Speech Signal Speech Signal Gammatone Filterbank Gammatone Filterbank Equal Loudness Pre-emphasis Equal Loudness Pre-emphasis (a) Intensity - Loudness Power Law LOG{ } (b) Inverse Discrete Fourier Transform Inverse Discrete Cosine Transform Auto Regressive Mod elling Smoothing Gamma-PLP Coefficients Gamma-Cepst Coefficients Block diagrams of two feature extraction paradigms. 32

33 Feature Evaluation Based on F-Ratio F-ratio is a measure of the feature effectiveness. It is the ratio of the between class variance (B) to the within class variance (W). For the i th feature in the j th class of K classes: F B i i Bi W i 1 K K j1 ( ij ) i 2 W i 1 K K j1 W ij 33

34 F-Ratio Based on HMM HMM satisfies the F-ratio conditions Features have Gaussian distribution. Diagonal covariance implies uncorrelated features For K states in each model and for H models we have: F ave 1 H H i1 F i 34

35 F-Ratio Characteristics F-ratio Mean F-ratio Q static delta delta-delta F-ratio of the between states procedure. The thick red line indicates the mean of the between states F-ratio. 35

36 Performance Evaluation F-ratio static delta delta-delta Q Classification properties based on F-ratio calculations of different feature extraction paradigms. 36

37 Feature Rank F-ratio MFCC Rank F-ratio GTCC Rank F-ratio GTPLP Rank F-ratio PLP Mean F- ratio

38 a MFCC13 b PLP13 c GTCC13 d GTPLP13 e Shows the states of the word three as detected by its four static features based CDHMMs. (a) Model MFCC13 is constructed from 13 static mel scale coefficients. (b) Model PLP13 is constructed from 13 static perceptual linear prediction coefficients. (c) Model GTCC13 is constructed from 13 static Gammatone cepstral coefficients. (d) Model GTPLP13 is constructed from 13 static Gammatone PLP coefficients. (e) The spectrogram of the input signal to envisage the frequency content of each state. 38

39 Classification Performance Absolute threshold recognition Margin Spoken words other than zero zero MFCC PLP GTPLP GTCC

40 Recognition Rate Performance DATASET-I DATASET-II Mel-cespt PLP Gamma-PLP Gamma-cepst DATASET-I : 10 digits DATASET-II : 31 words S/N ratio = 20 db 40

41 Conclusions Efficient auditory motivated technique is introduced. It is mainly based on Gammatone filterbank (GTF). GTF composed of non uniform bandpass filters imitating the frequency resolution of the cochlea. Two paradigms: Gamma-cepst and Gamma-PLP are investigated. Classification performance based on the F-ratio figure of merit has been investigated as it is a strong cue to the recognition performance. Gamma-cepst feature set outperforms the other feature sets. 41

42 42